I have an idea for agents (can be different models/wrappers) in interactive sessions (where they wait for user updates) to have a meeting. It's ridiculously simple but could be really effective. Have to prototype this tomorrow.
I have an idea for agents (can be different models/wrappers) in interactive sessions (where they wait for user updates) to have a meeting. It's ridiculously simple but could be really effective. Have to prototype this tomorrow.
TIL not to touch a lazy var from a bunch of detached tasks the hard way.
Sometimes one sets out to invent, and ended up discovering.
TIL about βPrompt frameworkβ. It sounds extremely similar to Design Patterns.
When the tool is insufficient, the pattern (framework) makes up for it.
Trying to reflect on why I was hugely impressed by Claude Codementia early on: itβs my bias towards TUI and its impressive ability to use CLI tools. The good impression faded quickly when I needed to solve real problems.
Are there any solutions for coding agents to chat with each other locally? Maybe through a CLI interface?
asking agents to read code for information feels like map reduce in my brain. I need to come up with questions as quickly as I can, and consolidate the small research reports later.
Stuck using MCPs is a special kind of hell.
Iβm experiencing software slop and Iβm not happy about it.
1M context window is just another way of saying expensive af.
There isnβt enough hours in a day.
Now instead of sending screenshots I just describe the visual tweak I want and Codex can go launch the app, navigate to where I mentioned, and look at it all by itself.
Asked Codex to learn to do end-to-end macOS app development using screenshots, OSScripts, pynput, etc. After it successfully played with some apps, I now asked it summarize everything it learned along the way to a skill for later use :)
Imagine how good will smith eating spaghetti will be in 100 years.
When people say programming is dead, what they mean is the economic value of rote pattern recognition + repetition is dead.
You can do it for fun and not profit.
You can profit without doing it.
Being good at it still profits you.
Mark my words.
I'm loving this workflow! I have a script that polls changes via the CLI, once it finds a todo item, it feeds its ID along side with a prompt to `codex exec`. The result is I can put todos in the queue and agents come and grab it. They mark the todo as finished as they do.
built a tool to help me work with multiple agents in parallel. I sit in the center coming up with ideas using a GUI and they poll for TODOs and claims them using a CLI. Tool's build and tested. Let's see how this goes!
looked into beads. TBH it seems like an overreaction to Claude Codementia.
I may genuinely need to make a TODO app for local coding agents. It'l have a CLI for agents and a GUI for me π€
I havenβt heard of those π
I may genuinely need to make a TODO app for local coding agents. It'l have a CLI for agents and a GUI for me π€
Doing more than 1 tasks at a time with agents still hurts my brain. I'll push through.
Working towards supporting embedded Swift for TOMLDecoder: github.com/dduan/TOMLDe...
An MCP is essentially a spec for a CLI app. Just ask your coding agent to build the CLI app and delete the token burner already :)
Modeling and 3D printing again. This brings me joy.
Update: Swift is now faster in both benchmarks: 23%, 6%
Running codex in a infinite fish shell loop that improves the code repeatedly measured by a benchmark reminds me of my auto character leveling script in World of Warcraft.
Runs 24/7, numbers go up π
Code is data. The way we generate data has changed. But LLMs can't generate the data by themselves.
while true; codex exec - < prompt.md; end
Found my old blog on archive.org. I was writing a lot about JavaScripts in 2012. I think I should bring those articles back for fun :)